14 Mar Seminar: “Conformal Prediction & Complex Data Analytics” by Dr. Matteo Fontana
March the 23th, 2023 at 3:00PM, on Teams: link.
In the latest years, scholars started focusing on how to develop statistical tool for the analysis of population of complex data, such as high-dimensional vector data and functions, but also more complex data objects such as functional time series or sets of graphs, either labelled or unlabelled. The present works adds to this literature by focusing on a strangely overlooked area, namely the formulation of prediction sets. By exploiting cutting edge techniques in the realm of machine learning, we propose a very powerful forecasting methodology, able to identify prediction regions in a very general sense, applicable to the great variety of possible data object the modern data scientist has to analyse. Our method, strongly based on Conformal Prediction, is model-free, achieves finite-sample validity, is computationally efficient and it identifies interpretable prediction sets, in the shape of a parallelotope. In the talk I will briefly present the basic ideas behind the methodology, its implementation to the functional and graph case, both labelled and unlabelled, as well as applications on real world data.
Matteo Fontana is a Project Officer at the Joint Research Centre of the European Commission, where is part of the Centre of Advanced Studies project “Computational Social Science for Policy”. From 2019 to 2021 he has been a Postdoctoral Researcher at the Modelling and Scientific Computing Lab of Politecnico di Milano, where he was involved in the development of an early warning system for geo-hazards in collaboration with the Italian Space Agency. He holds a PhD in Management Engineering from Politecnico di Milano, where he studied the application of novel statistical learning methodology to climate change economics research. He is a statistician/data scientist by training, he is mainly interested in applications of data science in economics, demography and migration studies. His main theoretical interests lie in the realm of nonparametric statistics (namely hypothesis testing and forecasting), as well as in the modelling of complex data objects.